Single-Channel Blind Source Separation for Singing Voice Detection: A Comparative Study

TL;DR

This paper introduces an unsupervised singing voice detection framework using single-channel blind source separation, compares three BASS methods, and evaluates their effectiveness with various features and machine learning models.

Contribution

It formalizes and compares three BASS approaches, extends the KAM method with a new training algorithm, and provides a comprehensive evaluation of singing voice detection methods.

Findings

BASS methods vary in separation and detection accuracy

The proposed KAM extension improves source-specific kernel computation

Unsupervised methods can achieve competitive singing voice detection performance

Abstract

We propose a novel unsupervised singing voice detection method which use single-channel Blind Audio Source Separation (BASS) algorithm as a preliminary step. To reach this goal, we investigate three promising BASS approaches which operate through a morphological filtering of the analyzed mixture spectrogram. The contributions of this paper are manyfold. First, the investigated BASS methods are reworded with the same formalism and we investigate their respective hyperparameters by numerical simulations. Second, we propose an extension of the KAM method for which we propose a novel training algorithm used to compute a source-specific kernel from a given isolated source signal. Second, the BASS methods are compared together in terms of source separation accuracy and in terms of singing voice detection accuracy when they are used in our new singing voice detection framework. Finally, we do an exhaustive singing voice detection evaluation for which we compare both supervised and unsupervised singing voice detection methods. Our comparison explores different combination of the proposed BASS methods with new features such as the new proposed KAM features and the scattering transform through a machine learning framework and also considers convolutional neural networks methods.